Subtasks Index

QA System Subtasks

An end-to-end QA system consists of many individual tasks. The key to improve the system is then to make progress in these tasks, so we are working on breaking things up, making intermediate datasets and measuring progress on these subtasks. (Of course, it also has caveats.)

Some of the tasks are common to all pipelines, others are specific to a particular architecture. We focus on the movies task in the subtask analysis at first; this e.g. means no unstructured data processing.

Movies Pipeline

Question Analysis - how to analyze the question and how to internally represent it?

Entity Linking - a particular question analysis aspect that deserves special attention due to its difficulty and importance.

Movies Knowledge Base - we do IR on top of this; does it have all we need, the way we need it?

Knowledge Base Relations - how to query the knowledge base and determine which data is relevant.

Answer Evidencing - extra analysis, especially for selection of the right type of answers

Answer Ranking

General Pipeline

Aside of the components of the movie pipeline, we also have:

Fulltext Information Retrieval - finding documents that answer the question

Answering Sentence Selection - finding answer-bearing sentences in question-related documents

Answer Extraction - extracting specific answer snippets from the answer-bearing sentences